Abstract:

Sketch-based image retrieval is an important research topic in the field of image processing. We present an improved feature extraction method for sketchbased image retrieval, which combines the Gaussian pyramid and local histogram of gradient (HOG) features. By using Gaussian pyramid, the image is decomposed into multiple resolution images on which the points of interest are extracted, and the points based multiscale HOG features are obtained. The visual word dictionary is generated from the multiscale HOG features, and finally feature vectors relative to visual words are formed. The images are ranked according to decreasing similarity which is calculated by the distance between the feature vectors of sketches and the images. Compared with the HOG algorithm at single scale and several other algorithms,  experimental results demonstrate the feasibility and effectiveness of the proposed method.

Key words: sketch-based image retrieval;Gaussian pyramid;multiscale HOG;shape feature